A Geometric Information Fused Loop Closing Method

Loop-closing methods play an important role in SLAM/SFM systems to eliminate the accumulative error.Existing loop-closing methods are based on the image retrieval which however easily fails to recognize the same scene viewed from different directions due to large perspective distortions.Recently,more and more SLAM/SFM systems are applied on small mobile platforms.These platforms shoot from different directions due to the applications.Hence monocular camera based loop-closing methods become more and more important.This paper proposes a novel loop-closing method to address this issue.The key idea is to fully explore the 3D structure reconstructed by the Visual SLAM/SFM system,to render virtual views from which the feature descriptors are extracted.The strategy of loop detection utilizes both real and virtual cameras: when loop detection fails between real cameras,virtualreal camera matching can also be used for loop searching,so that loops can be detected if virtual and real camera views are similar.Also,a visual SFM system is implemented to support located key frames and point clouds for loop closure.We conduct several tests including different scenarios on two selfcollected datasets.Experimental results show that the recall rate of our approach is increased by 29.6% comparing with the existing method,while both methods achieve a precision rate of 100%.Also,we discuss the performance of the loop closure using different types of descriptors,which gives much help on loop detection.